Mining belt conveyor intelligent monitoring method and system based on optical fiber sensing technology

Abstract

The invention belongs to the field of optical fiber sensing technology and artificial intelligence, and particularly relates to a mining belt conveyor intelligent monitoring method and system based on the optical fiber sensing technology. The monitoring method comprises the following steps: acquiring sound wave optical fiber sensing data and carrying out LogFBank preprocessing; constructing a sound wave optical fiber sensing state monitoring model based on the PANNs-CNN; performing feature extraction on the preprocessed sensing data by using a monitoring model; carrying out feature fusion on the extracted features by utilizing a deep learning convolutional network and a full-connection classification aggregation network structure; and outputting the data after feature fusion by using an output layer to obtain a state classification result of the belt conveyor. The belt conveyor intelligent monitoring system is constructed by using the monitoring method, and belt conveyor transportation is scientifically managed by using the monitoring system. According to the mining belt conveyor intelligent monitoring method and system based on the optical fiber sensing technology and the sound wave classification neural network, the reliability and stability of continuous monitoring of the state of the mining belt conveyor are improved, and the remarkable effect is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of optical fiber sensing, and in particular relates to an intelligent monitoring method and system for a mining belt conveyor based on optical fiber sensing technology. Background Art

[0002] As my country's coal mining industry expands in scale, mechanization and automation improve, safety issues caused by electromechanical equipment failures are becoming increasingly serious, and effective equipment safety monitoring is particularly important. The mining belt conveyor system is the key to the automated transportation of minerals in mines, and the current monitoring and early warning technology lags behind the requirements of safe production.

[0003] Acoustic fiber sensing technology (abbreviated as fiber sensing technology) is a technology that uses the propagation characteristics of light in optical fibers to monitor and measure sound waves. When sound waves act on optical fibers, they cause slight deformations in the optical fibers, thereby changing state parameters such as the phase, intensity or polarization of the light in the optical fibers. These changes can be monitored through corresponding demodulation technology and converted into electrical signals for processing and analysis, thereby obtaining information about the sound waves.

[0004] The integrity of mining belt conveyor rollers is the main source of safety hazards in mining belt conveyor transportation systems in coal mining. Due to long-term mineral transportation, the belt conveyor rollers are severely worn, which in turn causes belt damage, fire safety, and production accidents. The use of fiber optic sensing technology to evaluate the integrity of belt conveyor rollers has become a new method. Summary of the invention

[0005] In order to solve the technical problems existing in the prior art, the present invention uses optical fiber sensing technology and artificial intelligence audio classification algorithm to evaluate the integrity of the rollers of the mining belt conveyor transportation system, realize real-time monitoring of the belt conveyor transportation system, obtain the operating status data of the belt conveyor during transmission in real time, and warn of possible safety hazards in advance to ensure the safe and stable operation of the belt conveyor system. By collecting the acoustic fiber sensing data generated during the operation of the mining belt conveyor, the LogFBank feature acquisition method is used to better retain the high-dimensional features and feature correlation of the acoustic signal, and the deep learning PANNs-CNN model is used for feature learning and analysis. The vibration characteristics of different rollers are experimentally compared, and the safety and stability of the mining belt conveyor transportation system are comprehensively analyzed to monitor and alarm the roller abnormalities. The results show that the use of optical fiber sensing technology can effectively evaluate the operating status of the belt conveyor and the integrity of the rollers.

[0006] At the same time, the present invention designs and implements an intelligent monitoring system for a mining belt conveyor based on optical fiber sensing technology, which has the functions of real-time monitoring, fault warning, data recording and analysis of mining belt conveyor equipment.

[0007] The monitoring method and system can effectively solve the problems of low reliability and poor stability faced by optical fiber sensing technology in monitoring acoustic vibration events of mining belt conveyors. The monitoring method and system can solve the problems of location alarm and type identification of roller bearing damage and roller skin damage events that occur during the transportation of mining belt conveyor systems.

[0008] In the first aspect, the present invention provides an intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology, which adopts the following technical solution:

[0009] An intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology, wherein a sensing optical fiber is laid along the running route of the mining belt conveyor, comprising the following steps:

[0010] Obtain acoustic fiber optic sensor data from mining belt conveyors;

[0011] Perform LogFBank preprocessing on the acquired acoustic fiber sensing data;

[0012] Construct an acoustic fiber optic sensing state monitoring model based on PANNs-CNN;

[0013] The acoustic fiber sensing state monitoring model is used to extract features from the pre-processed acoustic fiber sensing data;

[0014] The extracted features are fused using a deep learning convolutional network and a fully connected classification aggregation network structure;

[0015] The data after feature fusion is outputted through the output layer to obtain the state classification result of the mining belt conveyor.

[0016] In the second aspect, an intelligent monitoring system for a mining belt conveyor based on optical fiber sensing technology is provided, which applies the aforementioned intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology, including:

[0017] The data sensing module is configured to obtain acoustic fiber optic sensing data of the mining belt conveyor;

[0018] A data transmission module is configured to transmit the acquired acoustic fiber sensing data in real time and stably;

[0019] The data preprocessing module is configured to perform LogFBank preprocessing on the acquired acoustic fiber sensing data;

[0020] The model building module is configured to build an acoustic fiber sensing condition monitoring model based on PANNs-CNN;

[0021] The feature extraction and fusion module is configured to extract features from real-time acoustic fiber data using an acoustic fiber sensing state monitoring model, and to fuse the extracted features using a deep learning convolutional network and a fully connected classification aggregation network structure;

[0022] The state classification module is configured to output the feature-fused data using the output layer to obtain the state classification result of the mining belt conveyor.

[0023] In summary, the present invention has the following beneficial technical effects:

[0024] The intelligent monitoring method and system for mining belt conveyors based on optical fiber sensing technology proposed in the present invention have obvious advantages over the existing abnormal events for abnormal monitoring and alarm tasks of belt conveyors, such as belt conveyor roller bearing damage events and belt conveyor roller skin damage events, and further solve the problems of low reliability and poor stability of optical fiber sensors in the pre-alarm technology of mining belt conveyors. At the same time, the present invention proposes an acoustic fiber data preprocessing method for converting acoustic fiber sensing data into audio classification data, and proposes an optimal solution strategy for optical fiber warning event classification tasks based on different acoustic fiber sensing data features, and proposes an intelligent perception solution for converting classification recognition results into state classification tasks in acoustic fiber sensing technology.

[0025] Comprehensive experimental analysis shows that the intelligent monitoring method and system for mining belt conveyors based on optical fiber sensing technology proposed in the present invention have relatively good performance in terms of accuracy, real-time performance, stability, etc., which provides important technical support for coal mining, improves the reliability and stability of real-time monitoring of abnormal audio vibration events of mining belt conveyors, and has achieved remarkable results. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of the steps of the intelligent monitoring method for a mining belt conveyor according to the first embodiment of the present invention;

[0027] Figure 2 This is a working principle diagram of the optical fiber sensing system based on Φ-OTDR technology according to the first embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of interference principle based on backscattered Rayleigh light according to the first embodiment of the present invention;

[0029] Figure 4 is a data set event data diagram of the acoustic vibration event data classification according to the first embodiment of the present invention;

[0030] Figure 5 This is the acoustic wave classification neural network model optical fiber sensing belt conveyor state monitoring algorithm of the first embodiment of the present invention, that is, the mDOS model working principle diagram;

[0031] Figure 6 This is an example diagram of the installation principle of the optical fiber pickup for a mining belt conveyor according to the first embodiment of the present invention;

[0032] Figure 7 1 is a comparative analysis diagram of the experimental confusion matrix of the first embodiment of the present invention;

[0033] Figure 8 It is an architecture diagram of an intelligent monitoring system for a mining belt conveyor based on optical fiber sensing technology according to a second embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0035] Embodiment 1

[0036] like Figure 1 As shown in the figure, the intelligent monitoring method for mining belt conveyors based on optical fiber sensing technology, laying sensor optical fibers along the walking route of the belt conveyor transportation system, analyzing the overall operation status of the belt conveyor, collecting and preprocessing the fault data of the belt conveyor rollers, analyzing and labeling the audio data of the roller faults, learning and training the audio classification algorithm features, deploying the intelligent monitoring system of the belt conveyor, deploying the optimal model and fault warning, and intelligent monitoring and management of mining belt conveyors. Specifically, the following steps are included:

[0037] Step 1, obtaining the acoustic fiber optic sensing data of the mining belt conveyor, including using a narrow linewidth laser as a light source and obtaining an audio vibration signal as acoustic fiber optic sensing data through an optical fiber pickup, wherein the acoustic fiber optic sensing data includes a variety of acoustic event data signals, including but not limited to: a belt conveyor stop signal, a belt conveyor normal operation signal, a roller bearing damage signal, and a roller skin damage signal.

[0038] The acquisition of acoustic fiber sensing data for mining belt conveyors relies on a DAS (Distributed Acoustic Sensing) fiber sensing system driven by Φ-OTDR (phase-sensitive optical time-domain reflectometer) technology, which includes a sensing fiber and a fiber pickup. The operating mechanism is shown in the following figure. Figure 2As shown, based on the narrow line width laser as the light source, the audio vibration signal is collected by the optical fiber pickup. It is mainly based on the principle of optical coherence monitoring. The light emitted by the laser transmitter is divided into two beams, one as the signal light to sense the disturbance information, and the other as the local oscillator light to contrast with the signal light. The two beams of light will superimpose the signal before entering the detector, and the audio vibration information is sensed by the change of the phase and amplitude of the light. The time difference between the optical signal and the incident pulse is t, c is the speed of light in a vacuum, and n is the effective refractive index inherent in the optical fiber core. The distance between the event information returned by the optical fiber and the system transceiver can be calculated. The formula is as follows:

[0039] S=c×t / 2n.

[0040] The way to achieve continuous acoustic wave monitoring of optical fiber is to use the Φ-OTDR optical fiber system to detect optical fiber vibration based on the interference principle of backscattered Rayleigh light. Figure 3 Take a section of optical fiber with a total length of L, and divide it into N equal sections, each of which has a length of ΔL = L / N. Backscattering Rayleigh can be regarded as the superposition of backscattered light vectors generated by M randomly distributed Rayleigh scattering elements in each section of optical fiber with a length of ΔL. For the mth such section, the electric field characteristics of its backscattered Rayleigh light can be intuitively displayed by the following formula:

[0041]

[0042] α represents the attenuation coefficient of the optical fiber, and L m represents the total length of the first m segments of the optical fiber. r and φ refer to the scattering coefficient and phase corresponding to the i-th backscattering event in the optical fiber segment of length ΔL, respectively. In Φ-OTDR technology, when a light pulse propagates in an optical fiber, the backscattered Rayleigh light generated by it will interfere within the pulse width range, thereby forming an interference spectrum. If we regard the scattering coefficient of each optical fiber segment as a constant value, uniformly represented by r, then the final detected optical signal intensity can be expressed as:

[0043]

[0044] The acquisition of acoustic fiber optic sensing data, including multiple acoustic vibration event data signals, refers to four groups of data obtained by testing under four working conditions, namely, the normal operation acoustic wave signal of the belt conveyor, the stop acoustic wave signal of the belt conveyor, the damage acoustic wave signal of the belt conveyor roller bearing, and the damage acoustic wave signal of the belt conveyor roller skin.

[0045] Step 2: Perform LogFBank preprocessing on the acquired acoustic fiber optic sensor data, including segmenting the acquired acoustic fiber optic sensor data and enhancing the acoustic fiber optic sensor data, including time stretching, time shifting, noise addition and pitch adjustment. At the same time, perform LogFBank data transformation to extract acoustic signal features, converting one-dimensional acoustic wave data into a two-dimensional Mel feature map to form a data set for acoustic vibration event data state classification.

[0046] This process covers three key steps: framing, pre-emphasis, and windowing. The length of the original WAV audio file stored in the computer hard disk is variable. First, it is divided into multiple fixed-length segments, namely frames. We usually set the duration of each frame to 10 to 30 milliseconds to ensure that each frame contains enough cycles while avoiding excessive signal changes. Due to differences in the sampling rate of digital audio, the dimension of each frame vector will also be different. In addition, considering that the high-frequency sound wave signal of belt conveyor noise is significantly attenuated during long-distance propagation, this results in relatively few high-frequency signal components after fast Fourier transform.

[0047] In order to improve the resolution of high-frequency signals, we will enhance the high-frequency part of each frame of speech signal. This process is called pre-emphasis of speech signal. In the previous framing process, the continuous speech signal was directly cut into multiple segments. However, this truncation operation may cause spectrum leakage problems. In order to eliminate the short-term signal discontinuity at the edges of each frame, we use a windowing operation to weight the framing with a window function of a certain length, that is, use a window function with a sliding function to divide the signal into several short time periods. The complete framing operation of the signal S(n) can be expressed by the following equation:

[0048] S ω (n) = ω(n)·S(n)

[0049] Where S ω (n) represents the nth frame signal after windowing, and ω(n) represents the window function.

[0050] Finally, a short-time Fourier transform (STFT) is performed to obtain the power spectrum, and then a set of triangular Mel filters are used to filter the power spectrum to obtain the Mel spectrum. The Log-Mel feature is obtained by performing a logarithmic operation on the Mel spectrum, and its expression is as follows:

[0051] LogMel(m,j)=Log(Sm(m,j)+ε)

[0052] Among them, S m(m, j) is the Mel spectrum of the sound signal, and ε is a very small constant used to avoid the zero value problem of the logarithm. The logarithmic operation takes into account the nonlinear characteristics of the human auditory system, making the extracted features more consistent with human sound perception.

[0053] Mel frequency cepstral coefficient MFCC feature is obtained by discrete cosine transform of Log-Mel feature, and its expression satisfies the following formula, where n is the index of MFCC coefficient and J is the number of MFCC coefficients:

[0054]

[0055] LogFBank feature extraction is a slight change based on the MFCC algorithm. It is based on the logarithmic transformation of the Bank feature extraction result as the final result. The calculation amount is relatively small compared to MFCC, and the correlation of the features is high. The one-dimensional audio data is converted into a two-dimensional Mel feature map through LogFBank transformation. The vertical axis of the feature map is the Mel feature, and the horizontal axis is the number of signal frames. The data set that constitutes the classification of sound wave vibration event data is shown in Figure 4 .

[0056] Step 3, construct an acoustic fiber optic sensor state monitoring model based on PANNs-CNN (PANNs, Pretrained Audio Neural Networks, pretrained audio neural network; CNN, Convolutional Neural Networks, convolutional neural network), including adding a multi-layer convolution module on the basis of the PANNs-CNN14 neural network model, so that the monitoring model can extract the abnormal acoustic vibration event characteristics of the belt conveyor in the acoustic fiber optic sensor data.

[0057] The PANNs-CNN-based acoustic fiber sensing state monitoring model, namely the mDOS model (which is the unified naming convention for the fiber sensing early warning model), adopts a deep fully convolutional neural network architecture, directly receives the spectrogram as input data, and incorporates the connectionist temporal classification technology at the output level of the model to solve the problem of different lengths of audio input sequences. This realizes the end-to-end training process of the entire system, which can directly convert the sound waveform signal into the corresponding category probability.

[0058] The LogFBank feature map is extracted from the audio signal after framing, pre-emphasis and windowing as the model input (as the overall input of the training data), and the feature map size is 1×1000×64, such as Figure 5, the acoustic fiber sensing state monitoring model includes: several convolution layers and pooling layers, and then after a convolution feature extraction operation, the convolution kernel size is 3×3, the feature map dimension becomes 64 dimensions, and the deformable convolution layer replaces the original convolution layer of the model, so that the model can better extract the vibration event features in the fiber sensing data. The deformable convolution adds an offset to the sampling position in the standard convolution operation (the offset is learned during the model training process), so that the convolution kernel can expand the field of view during the training process, so that it can be closer to the shape and size of the target event when sampling, and more robust. Assuming that the feature map we input is (B×C×H×W), B represents the number of feature maps processed at a time, C represents the number of channels, H×W represents the feature map size, and the convolution kernel size is (N×N), then the output feature map is calculated by the deformable convolution formula:

[0059]

[0060] p n Represents the offset of each point in the convolution kernel relative to the center point, w(p n ) represents the weight of the corresponding position on the convolution kernel, p0 can be regarded as each point on the output feature map, y(p0) represents the specific value of each point on the output feature map, x(p0+p n ) represents the specific value of the convolution sampling area on the input feature map corresponding to each point on the output feature map, Δp n Represents the offset of the deformable convolution.

[0061] After that, it passes through the batch normalization layer BN, and uses the ReLU activation function for curve transformation, and then passes through the pooling layer for feature extraction. After that, the above steps are repeated five times, and the size of the convolution kernel remains unchanged at 3×3 each time. The feature map dimension changes from 128 dimensions, 256 dimensions, 512 dimensions, and 1024 dimensions to 2048 dimensions. Finally, a fully connected network is added to transform the 2048 feature points into the desired output classification dimension to complete the model feature extraction. The model structure is as follows Figure 5 shown.

[0062] Step 4: Use the acoustic fiber optic sensor state monitoring model to extract features from the preprocessed acoustic fiber optic sensor data, including a convolution feature extraction operation on the preprocessed acoustic fiber optic sensor data with a convolution kernel size of 3×3, followed by a batch normalization layer BN, and using the ReLU activation function for curve transformation, followed by feature extraction through a pooling layer, so that the monitoring model focuses on the characteristics of the acoustic vibration event data signal.

[0063] Step 5: The extracted features are fused using a deep learning convolutional network and a fully connected classification aggregation network structure. After five convolution layer feature extractions with a convolution kernel size of 3×3, the feature map dimension changes from 128D to 256D to 512D to 1024D, and finally becomes 2048D. Finally, the output layer fully connected network is added to transform the 2048 feature points into the desired output classification dimension.

[0064] Step 6: Output the feature-fused data using the output layer to obtain the state classification result of the mining belt conveyor.

[0065] After outputting the Softmax activation function, the classification result of the monitoring model is obtained and the classification task is completed.

[0066] This first embodiment will apply the improved PANNs-CNN based acoustic wave classification neural network model optical fiber sensing belt conveyor state monitoring algorithm and name it mDOS model, deploy it on the Φ-OTDR optical fiber sensing system, and innovatively implement the optical fiber sensing monitoring algorithm for belt conveyor state that combines optical fiber sensing technology with artificial intelligence. The main innovations include the following:

[0067] (1) An acoustic fiber data preprocessing method is proposed to convert acoustic fiber sensing data into audio classification data.

[0068] (2) An improved PANNs-CNN-based acoustic fiber optic sensor belt conveyor condition monitoring algorithm mDOS model is proposed.

[0069] (3) Propose an optimal solution strategy for the task of identifying belt conveyor condition monitoring and early warning events based on different acoustic fiber optic sensor data characteristics.

[0070] (4) An intelligent perception system is proposed to convert audio classification results into state classification tasks in the belt conveyor state monitoring and early warning tasks using fiber optic sensing technology.

[0071] The following is a specific example of applying the above method:

[0072] This example uses a conveyor belt system in the right main tunnel of a coal mine. The belt system is 1.8 kilometers long and is located 800 meters underground. This example uses fiber optic monitoring and sensing technology to evaluate the overall operating status of the belt conveyor and the integrity of all rollers of the belt conveyor. By real-time and accurate restoration of the vibration information of the underground belt conveyor, possible roller damage points are found, and abnormal roller alarms are provided in time, providing a basis for the overall safety of the belt conveyor transportation system.

[0073] In this intelligent detection example of the belt conveyor system, we used a technology that combines the DAS fiber optic sensing system based on Φ-OTDR technology with a deep learning audio classification neural network model. The experimental setting of the pulse signal frequency is 10000Hz, the pulse width is 60ns, the loss coefficient is 1.4dB / km, and the total length of the optical fiber is 1000m. Excluding non-test point parts such as some redundancy in the original signal, the actual detection length is 600m, and 8 fiber optic pickup sampling points are set. This example also explores the sound restoration effect of different pickups. The first sampling point is set at 100 meters from the head of the belt conveyor, and then a sampling point is set every 10 meters. The use of optical fiber monitoring technology to detect coal mine belt conveyor systems has the following advantages:

[0074] (1) High precision: Fiber optic monitoring technology can achieve high-precision monitoring of physical quantities, with accuracy reaching sub-meter level.

[0075] (2) High sensitivity: Fiber optic monitoring technology can monitor tiny changes such as sound waves and vibrations.

[0076] (3) High reliability: Fiber optic monitoring technology can realize real-time monitoring of physical quantities in a large area with high reliability and stability.

[0077] (4) High safety: Fiber optic monitoring technology can realize real-time monitoring of physical quantities in dangerous areas, improving safety.

[0078] (5) Easy to install and maintain: Fiber optic monitoring technology can achieve monitoring through simple fiber optic layout and connection, which is easy to install and maintain.

[0079] In order to achieve real-time monitoring of the belt conveyor system, this example deploys a fiber optic sensing system. The system consists of a sensing fiber, a fiber optic pickup, and a distributed acoustic sensor host (DAS). The fiber optic pickup is used to collect acoustic wave data generated by the vibration of the roller during the operation of the belt conveyor. The sensing fiber is a data transmission device used to transmit the collected data to the acoustic sensor host (DAS) device for data processing and analysis.

[0080] In order to explore the changes in the acoustic audio signals that can be collected by the fiber optic pickup when the roller fails during the operation of the belt conveyor, we conducted the following simulation experiments, such as Figure 6 As shown, tapping experiments were performed at both ends of the fiber optic pickup when the roller was running and stopped, and the changes in the audio signal were observed and recorded.

[0081] We installed a fiber optic pickup about 100 meters away from the machine head to ensure that it is not affected by the vibration of the machine head. We tapped at both ends of the fiber optic pickup, and the tapping points were 1 meter, 3 meters, 5 meters, 7 meters, and 9 meters away from the fiber optic pickup. The tapping frequency was once every 1 second, and each position was tapped for 3 minutes. The data collected by the fiber optic pickup was collected, such as Figure 6 .

[0082] The experiment shows that the closer the knocking point is to the fiber optic pickup, the stronger the signal reaction is, and the farther the distance is, the weaker the signal is. When it exceeds 5m, the sound wave signal hardly changes. There is basically no difference in the signal obtained when the same fiber optic pickup is knocked at the same distance on the left and right sides, indicating that the sound wave data generated when the roller is running is relatively evenly propagated. In summary, the sensor fiber system is installed as follows. The sensor light is laid along the steel trough of the belt conveyor transportation system. It is known that the total length of the optical fiber is 1000m. Excluding the non-test point part such as some redundancy in the original signal, the actual detection length is 600m. 10 sampling points are set, of which 5 are pickup sampling points and 5 are fiber optic ring sampling points. The first sampling point is set at 100 meters from the head of the belt conveyor, and then a sampling point is set every 10 meters. This experiment can detect the overall operation status of the first 600 meters of the belt conveyor transportation system and the integrity status of the roller from 100 meters to 200 meters.

[0083] The implementation process of abnormal data collection of the intelligent monitoring system for mining belt conveyors is as follows: the overall 48-hour vibration signal of the entire belt conveyor when it is running and stopped is collected, which is used to train the model of normal operation of the equipment. At the same time, through investigation and analysis, the abnormal operation of the belt conveyor is mainly the abnormal roller, including the damage of the roller bearing and the damage of the roller skin. The experiment collects two kinds of damage state data of the roller, and each state data is not less than 3 hours, which is used for model alarm training.

[0084] In this example, the collected sound wave data of belt conveyor stop, belt conveyor normal operation, roller bearing damage, and roller skin damage are divided into several small time blocks of equipment audio data through segmentation. By determining the time interval between abnormal frequencies, the time block of t=10 seconds is finally determined. In order to avoid the problem of information loss in adjacent time blocks during the cropping process, the segmentation step is set to 8s, that is, there will be 2s of information overlap between two adjacent time blocks. Finally, all the noise data of the valid data set is determined, which contains 8000 audio segments, each audio segment is 10s long, and the sampling rate of the experimental setting is 10000HZ.

[0085] This example uses four data enhancement techniques, including time stretching, time shifting, noise addition, and pitch adjustment. Time stretching refers to changing the playback speed of the audio signal by adjusting the stretching factor v while keeping the pitch constant. This example sets a set of stretching parameter ranges v = {0.8, 2} for each audio data to avoid audio distortion. Time shifting is to slightly move the audio signal in the time domain while keeping the pitch unchanged. This example implements a random small time shift for each audio. Noise addition is to mix background noise into the original audio, such as Gaussian noise or ambient sound. This example chooses to add Gaussian white noise with a mean of 0 and a standard deviation of 1. Pitch adjustment is to change the pitch of the original audio while keeping the sound speed constant. Although this change does not affect the classification label of the fault feature, pitch adjustment as a data enhancement method has significant benefits for this study.

[0086] At the same time, this example uses Mel frequency to extract spectral feature information, and uses the paddleaudio.features.LogMelSpectrogram method in the Python third-party library paddleaudio to obtain the Log-Mel feature map. The feature map processed by this method is shown in the figure below.

[0087] This time, the deep learning Paddle framework was used for the experiment. The python version was 3.10. The fiber optic acoustic vibration sensor dataset was subjected to four audio classification tasks: roller stop, roller normal operation, roller bearing damage, and roller skin damage. The experimental platform was configured to use the Windows 10 operating system, 32G memory, Intel I7-12700 CPU, NVDIAGeForce RTX 3080TI GPU, and CUDA version 11.3. The main experimental parameter settings are shown in the following table.

[0088]

[0089]

[0090] The method used for effect evaluation is mainly used to evaluate the effect of the model on the test set.

[0091] In order to evaluate the effectiveness of the model alarm audio classification algorithm of the intelligent monitoring system of mining belt conveyors, this example uses precision, recall and accuracy as performance evaluation indicators. Precision focuses on evaluating the accuracy of the prediction results. It measures the proportion of samples that are truly positive among the instances predicted to be positive samples. Specifically, the calculation of precision takes into account two situations: one is to correctly identify the positive class as the positive class (true positive examples, TP), and the other is to incorrectly identify the negative class as the positive class (false positive examples, FP). Among them, TP is the number of sound clips whose predicted value is the same as the actual value, and FP is the number of sound clips in the prediction result that are incorrectly predicted to be a certain category. The calculation formula for precision is:

[0092]

[0093] The recall rate focuses on the proportion of positive examples in the original samples that are correctly predicted. It reflects how many samples that are actually positive examples are correctly identified as positive examples by the model. There are also two situations involved here: one is correctly predicting the original positive class as the positive class (true positive example, TP), and the other is incorrectly predicting the original positive class as the negative class (false negative example, FN). FN is the number of incorrectly predicted sound clips in a certain category of sound clips. The calculation formula for the recall rate is:

[0094]

[0095] The accuracy rate indicates the percentage of correctly predicted sound clips to the total number of sound clips in the test set. The accuracy rate is calculated as

[0096]

[0097] The precision, recall and average accuracy of the three models for four types of sound recognition are shown in the table below.

[0098] Table 1: PANNs-CNN model recognition rate display table

[0099]

[0100]

[0101] The results show that the recognition rate and average accuracy of the PANNS-CNN14 pre-trained model for the four categories of sound are higher than those of the other two models. Among them, the average accuracy of the PANNS-CNN6 model for the four categories of belt conveyor stop, belt conveyor normal operation, roller bearing damage, and roller skin damage is 76.25%, the average accuracy of the PANNS-CNN10 model for the four categories of belt conveyor stop, belt conveyor normal operation, roller bearing damage, and roller skin damage is 88%, and the average accuracy of the PANNS-CNN14 model for the four categories of belt conveyor stop, belt conveyor normal operation, roller bearing damage, and roller skin damage is 95%. When the three models are trained at the same time, the average prediction time of PANNS-CNN6 is the shortest, and the average prediction time of PANNS-CNN14 is the longest, but the average time for each audio segment of the same verification set does not exceed 0.2 seconds, which meets the needs of actual prediction. Therefore, PANNS-CNN14 is more suitable for the early warning task of the intelligent monitoring system of belt conveyors.

[0102] The confusion matrix is ​​mainly used to evaluate the performance of the model on the test set.

[0103] The confusion matrix is ​​an important tool for evaluating the prediction performance of classification problems. It summarizes the details of correct and incorrect classifications in the prediction results by counting, and divides these results into categories. In the field of fiber optic sensing pattern recognition, especially when evaluating the prediction performance of the PANNS-CNN model, the confusion matrix can clearly show which categories the algorithm is confused between, that is, the model mistakenly classifies samples of one category into another category. This matrix not only provides us with direct visual feedback on the performance of the algorithm, but also reveals possible biases or weaknesses in the model. More importantly, it deeply analyzes the specific types of errors that occur.

[0104] The comparison analysis of the confusion matrix of the PANNS-CNN14 experiment is shown in Figure 7 .

[0105] Through the confusion matrix, we can identify that the audio features of the normal operation of the belt conveyor are the most obvious and not easily confused with other features. The audio features generated by the damage of the roller skin are easily confused with other features. It can also be clearly heard at the experimental site that the audio features are similar to other audio features. The next step of the experiment can be to conduct a refined analysis of this specific type between specific categories that are easily confused, to make up for the deficiency of relying solely on classification accuracy as an evaluation indicator. Because although the classification accuracy can give an overview of the overall performance, it often masks the differences in prediction performance between different categories and the complexity of specific error patterns. However, PANNS-CNN14 can achieve a recognition rate of more than 90% for the audio of the normal operation of the belt conveyor, the audio of the stopped belt conveyor, the audio of the damaged roller bearing, and the audio of the damaged roller skin. We can believe that the audio classification model PANNS-CNN14 based on deep learning can effectively identify abnormal audio characteristics of the belt conveyor.

[0106] Embodiment 2

[0107] This embodiment 2 provides a mining belt conveyor intelligent monitoring system based on optical fiber sensing technology, which adopts the mining belt conveyor intelligent monitoring method based on optical fiber sensing technology described in embodiment 1. Figure 8 ,include:

[0108] (1) Data perception layer (data perception module). It consists of optical fiber sensors and microphones distributed at key locations of the belt conveyor. The optical fiber sensors can monitor multiple parameters such as temperature, vibration, and stress in real time, and the microphones can capture the sound changes during the operation of the belt conveyor system in real time.

[0109] (2) Data transmission layer (data transmission module). It includes sensor optical fiber, wireless network, etc., to ensure the real-time and stability of data transmission, and transmit the data collected by the data perception layer to the DAS data processing and algorithm alarm platform.

[0110] (3) Data processing layer (data preprocessing module, feature extraction and fusion module, state classification module). It includes data receiving function, signal processing function, data storage function, and algorithm alarm function. It is responsible for receiving data from the data transmission layer, preprocessing and algorithm processing the data, and then classifying and storing the processed data.

[0111] (4) Decision support layer (decision support module). It includes status assessment function, early warning decision function and maintenance guidance function. It is responsible for comprehensively evaluating the information of the data processing layer, judging the current operating status of the mining belt conveyor, and providing fault warnings for the mining belt conveyor. According to the operating requirements of the mining belt conveyor system, it provides possible troubleshooting suggestions and maintenance guidance.

[0112] Early warning and decision-making are carried out through the host computer program, and corresponding operations are performed according to the working conditions of the mining belt conveyor identified by the model.

[0113] (5) User interaction layer (user interaction module). This layer provides users with a visual interface, including real-time monitoring interface, historical data query, alarm information display, system maintenance and other functions, so that operators can easily understand and operate the system.

[0114] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology, wherein a sensing optical fiber is laid along the running line of the mining belt conveyor, characterized in that: The following steps are involved: Obtain acoustic fiber optic sensor data from mining belt conveyors; Perform LogFBank preprocessing on the acquired acoustic fiber sensing data; Construct an acoustic fiber optic sensing state monitoring model based on PANNs-CNN; The acoustic fiber sensing state monitoring model is used to extract features from the pre-processed acoustic fiber sensing data; The extracted features are fused using a deep learning convolutional network and a fully connected classification aggregation network structure; The data after feature fusion is outputted through the output layer to obtain the state classification result of the mining belt conveyor.

2. The intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology according to claim 1 is characterized in that: The method for obtaining the acoustic fiber optic sensing data of the mining belt conveyor includes using a narrow linewidth laser as a light source and obtaining an audio vibration signal as acoustic fiber optic sensing data through an optical fiber pickup. The acoustic fiber optic sensing data includes a variety of acoustic event data signals, including but not limited to: a belt conveyor stop signal, a belt conveyor normal operation signal, a roller bearing damage signal, and a roller skin damage signal.

3. The intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology according to claim 2 is characterized in that: The acquired acoustic fiber optic sensor data is subjected to LogFBank preprocessing, including segmenting the acquired acoustic fiber optic sensor data, and enhancing the acoustic fiber optic sensor data, including time stretching, time shifting, noise addition, and pitch adjustment. At the same time, LogFBank data transformation is performed to extract acoustic wave signal features, converting one-dimensional acoustic wave data into a two-dimensional Mel feature map to form a data set for acoustic wave vibration event data state classification.

4. The intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology according to claim 3 is characterized in that: The PANNs-CNN-based acoustic fiber sensing state monitoring model is constructed, including adding a multi-layer convolution module on the basis of the PANNs-CNN14 neural network model, so that the monitoring model can extract the abnormal acoustic vibration event features of the belt conveyor in the acoustic fiber sensing data.

5. The intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology according to claim 4 is characterized in that: The acoustic fiber optic sensing state monitoring model is used to extract features from the preprocessed acoustic fiber optic sensing data, including performing a convolution feature extraction operation on the preprocessed acoustic fiber optic sensing data, with a convolution kernel size of 3×3, then passing through a batch normalization layer BN, and using a ReLU activation function for curve transformation, and then performing feature extraction through a pooling layer, so that the monitoring model focuses on the characteristics of the acoustic vibration event data signal.

6. The intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology according to claim 5 is characterized in that: After five times of feature extraction through convolution layers with a convolution kernel size of 3×3, the feature map dimension changes from 128D to 256D to 512D to 1024D, and finally becomes 2048D. Finally, a fully connected network is added to the output layer to transform the 2048 feature points into the desired output classification dimension. After the output Softmax activation function, the classification result of the monitoring model is obtained, completing the classification task.

7. An intelligent monitoring system for mining belt conveyors based on optical fiber sensing technology, characterized in that: The intelligent monitoring method for a mining belt conveyor based on optical fiber sensing technology as claimed in claim 1 comprises: The data sensing module is configured to obtain acoustic fiber optic sensing data of the mining belt conveyor; A data transmission module is configured to transmit the acquired acoustic fiber sensing data in real time and stably; The data preprocessing module is configured to perform LogFBank preprocessing on the acquired acoustic fiber sensing data; The model building module is configured to build an acoustic fiber sensing condition monitoring model based on PANNs-CNN; The feature extraction and fusion module is configured to extract features from real-time acoustic fiber data using an acoustic fiber sensing state monitoring model, and to fuse the extracted features using a deep learning convolutional network and a fully connected classification aggregation network structure; The state classification module is configured to output the feature-fused data using the output layer to obtain the state classification result of the mining belt conveyor.

8. The intelligent monitoring system for mining belt conveyor based on optical fiber sensing technology according to claim 7 is characterized in that: Also includes: The decision support module is configured to determine the current operating status of the mining belt conveyor, and to provide fault warnings for the mining belt conveyor, troubleshooting suggestions and maintenance guidance; The user interaction module is configured with real-time monitoring interface, historical data query, alarm information display and system maintenance functions.

Images